(tokens, seq_lens=None)
| 145 | ) |
| 146 | |
| 147 | def replay(tokens, seq_lens=None): |
| 148 | self._prefill_inputs[0].copy_(tokens) |
| 149 | if seq_lens is not None: |
| 150 | self._prefill_inputs[1].k_seqinfo.seqlen.copy_(seq_lens) |
| 151 | |
| 152 | self._prefill_cuda_graph.replay() |
| 153 | torch.cuda.synchronize() |
| 154 | |
| 155 | return self._prefill_logits |
| 156 | |
| 157 | return replay |
| 158 |
nothing calls this directly
no outgoing calls
no test coverage detected